一种时间序列动态聚类的算法

针对时间序列传统静态聚类问题,提出了对时间序列进行动态聚类的方法。该方法首先提取时间序列的关键点集合,根据改进的FCM算法找到动态特征明显的时间序列,再利用提出的动态聚类算法确定此类时间序列在不同时间段的所属类别,在改进的FCM算法中采用兰氏距离可以使其对奇异值不敏感。实验结果反映出动态特征明显的时间序列类别随时间演化的特性,表明了方法的可行性和有效性。与已有算法相比,该方法揭示了时间序列的部分动态特征。该方法还可以运用于研究数据挖掘的其他问题。

Dynamic clustering algorithm for time series

This paper proposed a dynamic clustering algorithm for time series aiming at solving the shortcoming of traditional static clustering. Firstly, the method extracted the key point set of each time series, and then obtained the dynamic time series by using improved FCM algorithm. At last, detected the cluster of dynamic time series which belonged to each time segment based on the dynamic clustering algorithm. The adoption of L-W distance in FCM algorithm could avoid the shortcoming of sensitivity to singular value. The experimental results obtained by the proposal reflect the evolutional property that the clusters of the dynamic time series change over time, and show the validity and the feasibility of the method. Compared with existed algorithms, the proposed algorithm indicates the dynamic characteristic of time series when clustering them. This algorithm can also be applied to other problems in data mining.